CN109993224B - GEO satellite shape and attitude identification method based on deep learning and multi-core learning - Google Patents
GEO satellite shape and attitude identification method based on deep learning and multi-core learning Download PDFInfo
- Publication number
- CN109993224B CN109993224B CN201910239623.7A CN201910239623A CN109993224B CN 109993224 B CN109993224 B CN 109993224B CN 201910239623 A CN201910239623 A CN 201910239623A CN 109993224 B CN109993224 B CN 109993224B
- Authority
- CN
- China
- Prior art keywords
- sequence data
- ocs
- ocs sequence
- kernel
- convolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 69
- 238000013135 deep learning Methods 0.000 title claims abstract description 18
- 230000006870 function Effects 0.000 claims abstract description 65
- 239000013598 vector Substances 0.000 claims abstract description 60
- 238000012706 support-vector machine Methods 0.000 claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 30
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 21
- 238000013528 artificial neural network Methods 0.000 claims abstract description 12
- 238000005516 engineering process Methods 0.000 claims abstract description 11
- 125000004122 cyclic group Chemical group 0.000 claims abstract description 5
- 238000013507 mapping Methods 0.000 claims abstract description 4
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 230000004913 activation Effects 0.000 claims description 14
- 230000008569 process Effects 0.000 claims description 13
- 238000005070 sampling Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 9
- 230000003287 optical effect Effects 0.000 claims description 8
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000007500 overflow downdraw method Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000013145 classification model Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000005259 measurement Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000004088 simulation Methods 0.000 claims description 3
- 230000036544 posture Effects 0.000 description 20
- 238000010586 diagram Methods 0.000 description 8
- 230000000694 effects Effects 0.000 description 6
- 238000000605 extraction Methods 0.000 description 6
- 238000001914 filtration Methods 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 238000000513 principal component analysis Methods 0.000 description 5
- 238000012545 processing Methods 0.000 description 3
- 238000010200 validation analysis Methods 0.000 description 3
- 230000006399 behavior Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 238000002310 reflectometry Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 108091006503 SLC26A1 Proteins 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000000149 argon plasma sintering Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
- G01C21/24—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 specially adapted for cosmonautical navigation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Automation & Control Theory (AREA)
- Astronomy & Astrophysics (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a GEO satellite shape and attitude identification method based on deep learning and multi-core learning, which comprises the following steps: acquiring OCS sequence data of a GEO satellite for one year; preprocessing OCS sequence data; constructing a C-RNN model for automatically extracting characteristics of OCS sequence data based on a deep learning network, wherein the deep learning network comprises a cyclic neural network and a convolutional neural network; training a C-RNN model to obtain a plurality of feature vectors of OCS sequence data; based on the multi-kernel learning technology, a plurality of kernel functions are used for mapping different features, and a support vector machine is used for identifying the shape and the posture of the satellite. Under the condition of no need of prior information, the method combines deep learning based on the cyclic neural network and the convolutional neural network and the multi-core learning technology based on the support vector machine, utilizes a pure data driving mode, and uses OCS sequence data to automatically identify the shape and the posture of the GEO satellite.
Description
Technical Field
The invention belongs to the technical field of GEO satellite shape and attitude identification, and particularly relates to a GEO satellite shape and attitude identification method based on deep learning and multi-core learning.
Background
At present, the number of spatial objects is increasing, and spatial Situational Awareness (Space Situational Awareness) has become one of the international important research topics. Geosynchronous orbit (GEO) is an important space asset, verifying the state of all satellites and objects in GEO is very important for correctly assessing the GEO environment. Meanwhile, detailed spatial object characteristics (such as shape, posture and motion state) can be used for accurately predicting the track and behavior of the spatial object, and important information capability guarantee is provided for spatial situation perception. Optical observation is an important tool for acquiring information of a space object, the existing large number of optical telescopes are still the main methods for observing the GEO target in China, and more specifically, an analyst can extract the characteristics of the shape, the size, the posture, the reflectivity, the material and the like of the space object through photometric sequence data acquired by an optical observation system, and finally judge the behavior and the intention of the space target. Currently, an optical scattering cross section (OCS) is widely used to represent optical scattering characteristics of a spatial target, and the shape and posture of a GEO target can be effectively recognized through the visible light scattering characteristics. Since the OCS is only affected by the spatial target geometry, surface material, pose, and relative position of the sun-spatial target-observation station, but is independent of the observation distance and observation system, and the OCS sequence data and the photometric sequence data are convertible to each other, the OCS sequence data is well suited for feature recognition of the spatial target.
Currently, analysts have been able to manually identify the shape and attitude of satellites through photometric sequence data. Generally, the most common identification methods are physical models (e.g., inversion methods based on binner models) and filters (e.g., kalman filtering). The point pairing method based on the two-surface element model requires that a satellite body has Lambert reflection characteristics and a complex three-dimensional shape, and simultaneously requires that a satellite sailboard has mirror surfaces and Lambert reflection characteristics and is close to a plane structure. The Kalman filtering method takes characteristic parameters of a spatial target to be identified as unknown state parameters of a system for optimal estimation, the shape of the spatial target and the direction of an inertia axis of the target can be identified by utilizing lossless Kalman filtering, and the method has better identification performance and higher speed.
Although the traditional physical model and filtering method are fast, the traditional method needs much prior information and is greatly influenced by the quality of the model. Furthermore, the amount of data obtained from a large number of viewing devices is very large and manual identification of humans is no longer feasible.
The method based on the two-surface element model firstly has a lot of requirements on the structure of the target model, secondly, the method requires that the contribution of the reflectivity and the area of the target body to each observation point is the same, the body postures of the two observation points are necessarily the same, and in addition, the method requires that the time interval between the point pairs or the position interval of the observation stations is increased as much as possible so as to improve the identification effect of the model, so the method does not have universality and practicability at present.
The kalman filtering method needs to fuse phase angle data and luminosity sequence data, so certain priori knowledge is needed, and meanwhile, uncertainty of model parameters of a spatial target can cause a system error of kalman filtering estimation.
Disclosure of Invention
The invention aims to provide a GEO satellite shape and posture recognition method based on deep learning and multi-core learning, which is characterized in that under the condition of not needing prior information, the shape and posture of a GEO satellite are automatically recognized by using OCS sequence data in a pure data driving mode by combining deep learning based on a Recurrent Neural Network (RNN) and a Convolutional Neural Network (CNN) and a multi-core learning technology based on a Support Vector Machine (SVM).
In order to achieve the purpose, the invention is concretely realized by the following technical scheme:
the invention provides a GEO satellite shape and attitude identification method based on deep learning and multi-core learning, which comprises the following steps:
acquiring optical scattering cross section characteristic (OCS) sequence data of a GEO satellite;
step two, carrying out pretreatment on OCS sequence data;
constructing a C-RNN model for automatically extracting characteristics of OCS sequence data based on a deep learning network, wherein the deep learning network comprises a cyclic neural network and a convolutional neural network;
training a C-RNN model to obtain a plurality of feature vectors of OCS sequence data;
and fifthly, mapping different features by using a plurality of kernel functions based on the multi-kernel learning technology, and identifying the shape and the posture of the satellite by using a support vector machine.
The method for acquiring the OCS sequence data of the GEO satellite in the first step comprises the following steps:
and acquiring the OCS sequence data of the space target by one or more of numerical calculation, photometric sequence data obtained by actual observation and/or laboratory simulation measurement.
Further, the method for acquiring the OCS sequence data of the spatial target by using a numerical calculation method includes: obtaining OCS sequence data of a space target by using a BRDF model and an OCS calculation method;
the photometric sequence data obtained through actual observation can be converted into OCS sequence data, and the conversion relationship between the photometric sequence data and the OCS sequence data is as follows:
where m is the photometric data (stars, etc.) and r is the distance between the observation device and the spatial target.
In the second step, the method for preprocessing the OCS sequence data comprises the following steps:
according to observation geometric position relations corresponding to OCS sequence data obtained in different observation intervals, dividing the OCS sequence data obtained under different observation geometric position relations into different subsets, and setting the label of each subset as a subclass of the corresponding class; the observation geometric position relation is the position relation of the sun, the space target and the observation station.
In the third step, the method for constructing the C-RNN model for automatically extracting the characteristics of the OCS sequence data comprises the following steps:
the C-RNN model consists of an encoder, a decoder and a classifier;
the encoder is used for taking the OCS sequence data as input and generating a feature vector with fixed length as output;
the decoder reconstructs input OCS sequence data through the feature vectors generated by the encoder;
the classifier consists of three fully-connected layers using a ReLU activation function and an output layer using a sigmoid activation function; using the feature vector generated by the encoder as an input; the feature vectors are mapped to classes as outputs using a sigmoid function, obtaining a shape and a pose corresponding to the GEO satellite input OCS sequence data.
Further, the loss function of the C-RNN model is:
L＝MSE+loss
wherein MSE is a loss function for OCS sequence data reconstruction, and loss is a loss function for shape and posture classification process; the total loss is minimized by using back propagation and gradient descent methods during the C-RNN model training process.
Further, the encoder includes two 1-D convolutional layers with a rectifying linear unit (ReLU) activation function, which can be expressed as:
wherein x is an input value; the 1-D convolution uses a one-dimensional convolution core to perform convolution on the OCS sequence data, so that a feature vector of the OCS sequence data is extracted; the 1-D convolution is defined as:
wherein,
inputting OCS sequence data; n is the length of the sequence data; w and k represent the 1-D convolution kernel and the number of sliding steps, respectively;
is the output vector after convolution; n-nk+1，nkIs the size of the convolution kernel;
a dropout layer is arranged behind each convolution layer; a second convolutional layer followed by a flatten layer that converts multidimensional features into 1-D features; the resulting feature vector with the specified length is generated by passing the output of the flatten layer to two fully-connected layers, which use the ReLU activation function.
Further, in the decoder, two Gated Recursive Unit (GRU) networks are applied to complete the task of reconstructing the OCS input signal;
the decoder uses the difference value delta t between the characteristic vector and the sampling timeNAs input, N is the number of sampling points; copying the feature vector l times, wherein l is the set length of the output sequence of the decoder; the difference of the sampling time points is also copied for l times; the feature vectors are used to characterize the OCS sequence data, sampling time to determine where each point in the reconstructed sequence is located.
In the fourth step, the method for training the C-RNN model and acquiring a plurality of feature vectors of OCS sequence data comprises the following steps:
inputting OCS sequence data with the length of 200 into a C-RNN model; the C-RNN model needs to be trained for 2 times, and a convolution kernel with the size of 5 is applied to two CNN layers for the first time; applying a convolution kernel with the size of 3 to the two CNN layers for the second time, respectively setting the iteration times and the batch size to 2000 and 1000, and storing the characteristic vector output by the C-RNN model after each training is finished;
wherein, two CNN layers in the model are respectively provided with 60 filters and 100 filters; the output of the encoder is an embedded feature vector of size 64;
the decoder uses a 2-layer unit size of 100 bi-directional GRU layer; the input to the decoder is a length 65 feature vector comprising the output of the encoder (length 64) and the sample time difference between OCS sequence data points (length 1);
the classifier processes the feature vectors generated by the encoder and provides a classification result of each OCS sequence data;
the Adam optimizer is used for network optimization, and the learning rate is 1 multiplied by 10-3(ii) a The dropout ratio of each dropout layer is set to 0.25.
In the fifth step, the method for recognizing the shape and the posture of the satellite includes:
based on the MKL technology, feature vectors generated by using C-RNN models of different convolution kernels are used as input data of a Support Vector Machine (SVM), basic kernel functions are combined through a multi-kernel linear combination method, a final kernel function formed by linearly combining multiple kernels is used as a kernel function of the SVM, and classification is carried out by using the SVM. The multi-core linear combination can be described as follows:
wherein X, z ∈ X,
as a result of the characteristic space,
for the ith normalized fundamental kernel function, K (x, z) represents the final kernel function formed by linearly combining n fundamental kernel functions, βiRepresents the ith coefficient; the basic kernel function is a polynomial kernel function, which can be expressed as:
in the formula: x, z ∈ X,
is a feature space; r and d are respectively constant and polynomial order; wherein, MKL is a multi-feature fusion method, and SVM is a classification model. Currently, the commonly used multi-core learning (MKL) is a multi-feature fusion method based on SVM. Generally, SVMs are single-kernel, and it is difficult to select the most suitable kernel function and corresponding parameters to obtain the best classification effect. And the MKL adopts different kernels according to different characteristics, different weights are distributed to the different kernels, then the weight of each kernel is trained, and the optimal combination of kernel functions is selected to complete a classification task.
The invention has the beneficial effects that:
the method utilizes the neural network to automatically extract the OCS sequence data of the GEO satellite; the C-RNN model provided by the invention comprises an encoder consisting of CNN, a decoder consisting of RNN and a classifier consisting of a fully-connected neural network; the main function of the classifier in the C-RNN architecture is to maximize the distance between features, rather than performing classification; the shape and the attitude of the GEO satellite are recognized by a multi-core learning technology and an SVM; the multi-core learning mode of the characteristic is realized by linearly combining a plurality of polynomial cores. The method can automatically extract the characteristics of the obtained OCS sequence data, thereby saving a large amount of labor cost; the extracted features contain richer information of OCS sequence data, and the classification effect can be improved during classification; different characteristics can be fused by using multi-core learning, so that the classification result is more accurate.
Drawings
FIG. 1 is a schematic diagram showing the structure of the C-RNN model.
FIG. 2 shows a schematic diagram of MKL.
Fig. 3a to 3e are schematic diagrams of 5 satellite models.
Fig. 4a to 4e are schematic diagrams illustrating the OCS sequence data reconstruction results of 5 satellites in the attitude 2.
FIG. 5 is a diagram illustrating the results of classifying OCS sequence data using a C-RNN structure of a convolution kernel of size 5.
FIG. 6 is a diagram illustrating the results of classifying OCS sequence data using a C-RNN structure of a convolution kernel of size 3.
Fig. 7 is a diagram illustrating the classification results of the test OCS sequence data by the MKL based support vector machine.
FIG. 8 is a graph showing training and validation losses for ENDECLA-CR, ENDE-RR.
Detailed Description
In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Example one
The embodiment of the invention provides a GEO satellite shape and attitude identification method based on deep learning and multi-core learning, which comprises the following steps:
the method comprises the steps of firstly, obtaining optical scattering cross section characteristic (OCS) sequence data of the GEO satellite.
The method for acquiring the OCS sequence data of the GEO satellite comprises the following steps: and acquiring the OCS sequence data of the space target by one or more of numerical calculation, photometric sequence data obtained by actual observation and/or laboratory simulation measurement. The method for acquiring the OCS sequence data of the space target in a numerical calculation mode comprises the following steps: obtaining OCS sequence data of a space target by using a BRDF model and an OCS calculation method; the photometric sequence data obtained through actual observation can be converted into OCS sequence data, and the conversion relationship between the photometric sequence data and the OCS sequence data is as follows:
where m is the photometric data (stars, etc.) and r is the distance between the observation device and the spatial target.
Step two, carrying out pretreatment on OCS sequence data; the method comprises the following steps:
according to observation geometric position relations corresponding to OCS sequence data obtained in different observation intervals, dividing the OCS sequence data obtained under different observation geometric position relations into different subsets, and setting the label of each subset as a subclass of the corresponding class; the observation geometric position relation is the position relation of the sun, the space target and the observation station. For example, if the OCS sequence data of the target T in the category C corresponds to n primary observation geometric positional relationships (n ≦ 5), the OCS sequence data of the target T in the category C will be divided into n sub-categories, and the label of each sub-category is set to C1, C2. Thus, OCS sequence data for one year will be divided into NCN classes, wherein NCIs the number of categories. The processing method can improve the training effect in the fourth step. In addition, each segment of OCS sequences is processed into sequence data of length 200 for identification classification.
Constructing a C-RNN model for automatically extracting characteristics of OCS sequence data based on a deep learning network, wherein the deep learning network comprises a cyclic neural network and a convolutional neural network;
in order to recognize the shape and orientation of the GEO satellite using the OCS data, it is necessary to compress the entire OCS sequence data into a feature vector. The C-RNN feature extraction model is shown in FIG. 1, where M and D represent the dimensions of CNN convolutional layers, the size of each convolutional layer is determined by the given convolutional kernel size, the number of filters and the length of input data, and the dropout layer is omitted.
The C-RNN model consists of an encoder, a decoder and a classifier; the encoder consists mainly of CNN, which takes the OCS sequence as input and generates a fixed length feature vector as output. The decoder reconstructs input OCS sequence data through the feature vectors generated by the encoder; the classifier consists of three fully-connected layers using a ReLU activation function and an output layer using a sigmoid activation function; using the feature vector generated by the encoder as an input; the feature vectors are mapped to classes as outputs using a sigmoid function, obtaining a shape and a pose corresponding to the GEO satellite input OCS sequence data.
Since the proposed model contains two outputs, two loss functions need to be defined. The loss function for the C-RNN model is:
L＝MSE+loss
wherein MSE is a loss function for OCS sequence data reconstruction, and loss is a loss function for shape and posture classification process; the total loss is minimized by using back propagation and gradient descent methods during the C-RNN model training process. The loss function used for the reconstruction of the OCS sequence is the Mean Square Error (MSE). The MSE may be given by:
in the formula:
for the ith sequence of the OCS,
for the ith reconstructed sequence, wiFor the weight coefficients, N represents the length of the output sequence and N represents the total number of input OCS sequences.
Binary cross entropy is used as a loss function for the shape and pose classification process. It should be noted that when binary cross entropy is used as the loss function, the tag of the sequence data needs to be binarized, and the output of the classifier is also a vector consisting of 0 and 1. The binary cross entropy is given by:
in the formula:
and
after binarization, the ith numerical value in the label and the ith value, n, in the category vector output by the classifier are respectivelycRepresenting the length of the classifier output vector.
The encoder consists mainly of CNN, which takes the OCS sequence as input and generates a fixed length feature vector as output. Including two 1-D convolutional layers with a rectifying linear unit (ReLU) activation function, which can be expressed as:
wherein x is an input value; the 1-D convolution uses a one-dimensional convolution core to perform convolution on the OCS sequence data, so that a feature vector of the OCS sequence data is extracted; the 1-D convolution is defined as:
wherein,
inputting OCS sequence data; n is the length of the sequence data; w and k represent the 1-D convolution kernel and the number of sliding steps, respectively;
is the output vector after convolution; n-nk+1，nkIs the size of the convolution kernel; in order to prevent overfitting of the neural network, a dropout layer follows each convolution layer; a second convolutional layer followed by a flatten layer that converts multidimensional features into 1-D features; the resulting feature vector with the specified length is generated by passing the output of the flatten layer to two fully-connected layers, which use the ReLU activation function.
In the decoder, two Gated Recursive Unit (GRU) networks are applied to complete the task of reconstructing the OCS input signal; GRU is a variant of RNN that effectively addresses the issue of long term dependence of RNN and performs better than standard RNN in the task of processing time series data.
The decoder uses the feature vector and the difference Δ t between the sampling timesNAs input, N is the number of sampling points; copying the feature vector l times, wherein l is the set length of the output sequence of the decoder; the difference of the sampling time points is also copied for l times; the feature vectors are used to characterize the OCS sequence data, sampling time to determine where each point in the reconstructed sequence is located. The C-RNN model can be used to process non-uniformly sampled time series data by appending the time difference between sample points as input.
The classifier in the C-RNN model can be used for processing multi-label classification problems, and the shape and the posture of the GEO target can be directly classified by using the classifier. However, in the feature extraction C-RNN model, the classifier is mainly used to maximize the distance between feature vectors corresponding to different targets and different poses, so that the feature information extracted by the encoder is richer and more accurate. The classifier consists of three fully connected layers using the ReLU activation function and one output layer using the sigmoid activation function. The feature vectors produced by the encoder are the input to the classifier. The output layer maps features to classes using a sigmoid function. By the classifier, the shape and orientation corresponding to the input OCS sequence data of the GEO target can be directly obtained.
Training a C-RNN model to obtain a plurality of feature vectors of OCS sequence data; the method comprises the following steps:
the input layer of the C-RNN model will process OCS sequence data of length 200. Inputting the preprocessed OCS sequence data with the length of 200 into a C-RNN model; the C-RNN model needs to be trained for 2 times, and a convolution kernel with the size of 5 is applied to two CNN layers for the first time; applying a convolution kernel with the size of 3 to the two CNN layers for the second time, respectively setting the iteration times and the batch size to 2000 and 1000, and storing the characteristic vector output by the C-RNN model after each training is finished;
wherein, two CNN layers in the model are respectively provided with 60 filters and 100 filters; the output of the encoder is an embedded feature vector of size 64;
the decoder uses a 2-layer unit size of 100 bi-directional GRU layer; the input to the decoder is a length 65 feature vector comprising the output of the encoder (length 64) and the sample time difference between OCS sequence data points (length 1);
because the classifier needs to process the multi-label classification problem, the label of each OCS sequence data needs to be binarized, and the classifier processes the feature vector generated by the encoder and provides the classification result of each OCS sequence data;
the Adam optimizer is used for network optimization, and the learning rate is 1 multiplied by 10-3(ii) a The dropout ratio of each dropout layer is set to 0.25.
And fifthly, mapping different features by using a plurality of kernel functions based on the multi-kernel learning technology, and identifying the shape and the posture of the satellite by using a support vector machine. The method comprises the following steps:
based on the MKL technology, the feature vectors generated by using C-RNN models with different convolution kernels are used as input data of a Support Vector Machine (SVM), and a basic kernel function is combined through a multi-kernel linear combination method, wherein the multi-kernel linear combination can be described as follows:
wherein X, z ∈ X,
as a result of the characteristic space,
for the ith normalized fundamental kernel function, K (x, z) represents the final kernel function formed by linearly combining n fundamental kernel functions, βiRepresents the ith coefficient; the basic kernel function is a polynomial kernel function, which can be expressed as:
in the formula: x, z ∈ X,
is a feature space; r and d are respectively constant and polynomial order; wherein, MKL is a multi-feature fusion method, and SVM is a classification model. And applying the final kernel function to the SVM classifier so as to determine the shape and the posture of the GEO target. The schematic diagram of multi-core learning is shown in fig. 2.
The invention has the beneficial effects that:
the method utilizes the neural network to automatically extract the OCS sequence data of the GEO satellite; the C-RNN model provided by the invention comprises an encoder consisting of CNN, a decoder consisting of RNN and a classifier consisting of a fully-connected neural network; the main function of the classifier in the C-RNN architecture is to maximize the distance between features, rather than performing classification; the shape and the attitude of the GEO satellite are recognized by a multi-core learning technology and an SVM; the multi-core learning mode of the characteristic is realized by linearly combining a plurality of polynomial cores. The method can automatically extract the characteristics of the obtained OCS sequence data, thereby saving a large amount of labor cost; the extracted features contain richer information of OCS sequence data, and the classification effect can be improved during classification; different characteristics can be fused by using multi-core learning, so that the classification result is more accurate.
To verify the effect of the present invention, the OCS sequence data of 5 different satellites in 3 different attitudes (three-axis stable mode) for one year is calculated by using the china lijiang astronomical stage (25.6 ° N,101.1 ° E, 2.465Km), and the model structures of the five spatial objects are shown in fig. 3 a-3E.
The obtained OCS sequence data can be classified into 15 major categories, which are: firstly, aiming at1 and attitude 1; an object 1 and a posture 2; target 1, attitude 3; target 2 and posture 1; fifthly, obtaining a target 2 and an attitude 2; sixthly, target 2 and posture 3; target-quietness 3, pose 1; and object 3, pose 2; a self-supporting target 3, a posture 3; objective 4, pose 1; target 4, attitude 2; a left target 4, a posture 3; a selection target 5, a gesture 1; first-loop target 5, attitude 2; original target 5 and posture 3. The three postures are respectively: 1) the x axis points to the satellite speed direction, the z axis points to the ground, and the x, y and z axes are orthogonal to each other and meet the right-hand rule; 2) the y axis points to the satellite speed direction, the z axis is along the solar panel direction, and the x, y and z axes are orthogonal to each other and meet the right-hand rule; 3) the z-axis points in the direction of the satellite velocity, the x-axis points in the earth's geocentric, and the x, y, z-axes are orthogonal to each other and satisfy the right hand rule.
In step 1, 6,915 (15 total) OCS sequence data for five GEO satellites in three poses, 1,383 each, are obtained, each corresponding to a different observation interval. OCS sequence data having a length of less than 200 is deleted, so that the available data amount to 5,505, 1,101 satellites each, and 367 satellites each. According to the observation geometric position relationship of the observation interval, dividing the OCS sequence data corresponding to each satellite into 5 sub-categories (namely 5 main observation geometric position relationships): the 1 st subclass has 129 OCS sequence data; the second subclass has 420 OCS sequence data; subclass 3 has a total of 141 OCS sequence data; subclass 4 has 399 OCS sequence data; there are 12 OCS sequence data in the 5 th subclass. To make numerically calculated OCS data more reliable, random error values satisfying a gaussian distribution are added to the original OCS data. Approximately 70% of the data in the OCS sequence data set was extracted as the training set, and the rest of the data were used as the validation set and the test set, respectively. After data pre-processing, the number of OCS sequences in the training set is approximately 12,500.
Inputting training data into the established model, and obtaining the trained model after 2000 times of training iterations. Approximately 2,160 OCS sequence data were used as test data, which were divided into 15 categories of 144 pieces of data each (each category representing one satellite in one attitude). The trained model is used to reconstruct and classify the test data, and the result of reconstructing the original test OCS sequence data under the attitude 2 for 5 satellites by the decoder in the first training is shown in fig. 4a to 4 e. In the figure, "Sat 1" represents "satellite 1" and "Attitude TWO" represents the 2 nd Attitude of the 3 satellite attitudes.
After two times of training, the classification result of the C-RNN architecture classifier is shown in fig. 5 and 6, where "T0 a 1" represents satellite 1, attitude 1, and the value on the diagonal is the number of correct classifications. When the encoder uses a convolution kernel with the size of 5, the classification accuracy of the classifier reaches 91.9%; when a convolution kernel of size 3 was used, the classification accuracy of the classifier was 83.3%. The classification results are different because the encoder will produce different feature vectors due to the different sizes of the convolution kernels. With the unchanged model structure, the encoder with a convolution kernel of size 5 can obtain features from the encoder that can better represent the OCS sequence data for three attitudes of five satellites.
To perform classification using multi-kernel learning, R in the basic polynomial kernel is set to 0. According to the obtained 2 characteristics, several polynomial kernels are constructed in sequence, wherein the order of each polynomial kernel function is d ═ 1, 2, 3, …,10 respectively. And (3) taking a Support Vector Machine (SVM) as a classifier, taking the final combined kernel K as a kernel function of the SVM, processing a training set and training. After the training of the classifier is finished, the optimal penalty value C of the SVM is 1000, and the linear combination coefficients of the kernel function are all 0. The results of the MKL classification based on SVM are shown in FIG. 7. The classification accuracy rate obtained by multi-core learning reaches 99.58%.
In order to evaluate the feature selection performance of the proposed C-RNN architecture (the proposed C-RNN architecture is referred to as ENDECLA-CR), several network models were first constructed to compare with the proposed C-RNN architecture. The first network model is a model that removes classifiers in the C-RNN (this model is denoted ENDE-CR); the second network model is to replace the two convolutional layers of the encoder in the ENDE-CR with 2 GRU layers of size 100 (this model is called ENDE-RR). The model parameters and training parameters used by these two models are consistent with the parameters used by the proposed C-RNN model. Training data used to train the C-RNN model is used as input training data for the ENDE-CR and ENDE-RR models. The training loss and validation loss for ENDECLA-CR, ENDE-RR after training is complete are shown in FIG. 8.
Meanwhile, table 1 lists the classification accuracy of the present invention and the other 6 feature extraction models for comparison, showing the superior recognition performance of the present invention. The remaining 6 feature extraction models include Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Dictionary Learning (DL) and ENDE-CR, ENDE-RR, and a simple deep neural network (the encoder and decoder of ENDE-RR are replaced with a simple multi-layer fully-connected structure, called SDNN). The SDNN model contains 1 input layer (200 cells), 4 fully connected layers (500 cells per hidden layer) and 1 output layer (200 cells). The number of training iterations for SDNN was set to 2000, the loss function used MSE, and the optimizer used Adam.
The training set and test set used in the comparative model are the same as those used in the C-RNN model. And respectively inputting the features extracted by the trained 7 models into a support vector machine with a linear kernel for classification. The reason for using this classifier is that it does not map the input features into a high-dimensional space or perform other transformations, which can guarantee the validity of the comparison result.
The recognition performance of the model is evaluated by the average accuracy MAP of each model class, which is the average prediction accuracy of all classes and can be expressed as:
in the formula: n is the total number of classes, precisioncClassification accuracy of class c, xTcThe number of correctly predicted OCS sequences of the category c is Mc, and the total number of OCS sequences of the category c is identified as Mc. K in the table represents the size of the convolution kernel used by the C-RNN structure.
TABLE 1
As can be seen from the classification performance of 7 models in Table 1, the recognition performance of the PCA model and the LDA model respectively reaches 61% and 76%; compared with PCA and LDA, DL has better recognition performance, and the recognition accuracy rate is 73%; the identification performance of the SDNN is superior to that of the traditional feature extraction methods such as PCA, LDA and DL, and the identification accuracy rate reaches 84.5%; ENDE-CR and ENDE-RR have great advantages in the aspect of feature extraction, and the recognition accuracy rate respectively reaches 95.8 percent and 83 percent. From the results, it can be seen that the recognition accuracy of the proposed C-RNN structure is the best. When the sizes of the convolution kernels are 3 and 5, the recognition accuracy exceeds 98%. The classifier in the C-RNN increases the distance between different classes of features, so that the feature vectors generated by the encoder can better represent OCS sequence data, and the proposed C-RNN architecture has good classification performance.
The above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments.
It should be noted that, for simplicity of description, the above-mentioned method embodiments are described as a series of acts or combination of acts, but those skilled in the art will recognize that the present invention is not limited by the order of acts, as some steps may occur in other orders or concurrently in accordance with the invention. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and modules referred to are not necessarily required by the invention.
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.
Claims (9)
1. A GEO satellite shape and attitude identification method based on deep learning and multi-core learning is characterized by comprising the following steps:
acquiring optical scattering cross section characteristic (OCS) sequence data of a GEO satellite;
step two, carrying out pretreatment on OCS sequence data;
constructing a C-RNN model for automatically extracting characteristics of OCS sequence data based on a deep learning network, wherein the deep learning network comprises a cyclic neural network and a convolutional neural network;
training a C-RNN model to obtain a plurality of feature vectors of OCS sequence data;
mapping different feature vectors by using a plurality of kernel functions based on a multi-kernel learning technology, and identifying the shape and the posture of the satellite by using a support vector machine;
in the third step, the method for constructing the C-RNN model for automatically extracting the characteristics of the OCS sequence data comprises the following steps: the C-RNN model consists of an encoder, a decoder and a classifier; the encoder is used for taking the OCS sequence data as input and generating a feature vector with fixed length as output; the decoder reconstructs input OCS sequence data through the feature vectors generated by the encoder; the classifier consists of three fully-connected layers using a ReLU activation function and an output layer using a sigmoid activation function; using the feature vector generated by the encoder as an input; feature vectors are mapped to classes as output using a sigmoid function, obtaining the distance between the maximized features.
2. The method of claim 1, wherein the method of acquiring the OCS sequence data of the GEO satellite in the first step comprises:
and acquiring the OCS sequence data of the space target by one or more of numerical calculation, photometric sequence data obtained by actual observation and/or laboratory simulation measurement.
3. The method of claim 2, wherein the method of obtaining the OCS sequence data of the spatial object by means of numerical calculation comprises: obtaining OCS sequence data of a space target by using a BRDF model and an OCS calculation method;
the photometric sequence data obtained through actual observation can be converted into OCS sequence data, and the conversion relationship between the photometric sequence data and the OCS sequence data is as follows:
where m is the photometric data and r is the distance of the observation device from the spatial target.
4. The method of claim 1, wherein in the second step, the method for preprocessing the OCS sequence data comprises:
according to observation geometric position relations corresponding to OCS sequence data obtained in different observation intervals, dividing the OCS sequence data obtained under different observation geometric position relations into different subsets, and setting the label of each subset as a subclass of the corresponding class; the observation geometric position relation is the position relation of the sun, the space target and the observation station.
5. The method of claim 1, wherein the loss function of the C-RNN model is:
L＝MSE+loss
wherein MSE is a loss function for OCS sequence data reconstruction, and loss is a loss function for shape and posture classification process; the total loss is minimized by using back propagation and gradient descent methods during the C-RNN model training process.
6. The method of claim 1, wherein the encoder includes two 1-D convolutional layers with a rectifying linear unit (ReLU) activation function, the ReLU function being expressed as:
wherein x is an input value; the 1-D convolution uses a one-dimensional convolution core to perform convolution on the OCS sequence data, so that a feature vector of the OCS sequence data is extracted; the 1-D convolution is defined as:
wherein,
inputting OCS sequence data; n is the length of the sequence data; w and k represent the 1-D convolution kernel and the number of sliding steps, respectively;
is the output vector after convolution; p ═ n-nk+1，nkIs the size of the convolution kernel;
a dropout layer is arranged behind each convolution layer; a second convolutional layer followed by a flatten layer that converts multidimensional features into 1-D features; the resulting feature vector with the specified length is generated by passing the output of the flatten layer to two fully-connected layers, which use the ReLU activation function.
7. The method of claim 1, wherein in the decoder, two Gated Recursive Unit (GRU) networks are applied to accomplish the task of reconstructing the OCS input signal;
the decoder uses the difference Δ t between the feature vector and the sampling timeNAs input, N is the number of sampling points; copying the feature vector l times, wherein l is the set length of the output sequence of the decoder; the difference of the sampling time points is also copied for l times; the feature vectors are used to characterize the OCS sequence data, sampling time to determine where each point in the reconstructed sequence is located.
8. The method of any one of claims 1-7, wherein in step four, the method for training the C-RNN model to obtain a plurality of feature vectors of OCS sequence data comprises:
inputting OCS sequence data with the length of 200 into a C-RNN model; the C-RNN model needs to be trained for 2 times, and a convolution kernel with the size of 5 is applied to two CNN layers for the first time; applying a convolution kernel with the size of 3 to the two CNN layers for the second time, respectively setting the iteration times and the batch size to 2000 and 1000, and storing the characteristic vector output by the C-RNN model after each training is finished;
wherein, two CNN layers in the model are respectively provided with 60 filters and 100 filters; the output of the encoder is an embedded feature vector of size 64;
the decoder uses a 2-layer unit size of 100 bi-directional GRU layer; the input to the decoder is a length 65 eigenvector, which includes the length 64 encoder output and the sample time difference between the length 1 OCS sequence data points;
the classifier processes the feature vectors generated by the encoder and provides a classification result of each OCS sequence data;
the Adam optimizer is used for network optimization, and the learning rate is 1 multiplied by 10-3(ii) a The dropout ratio of each dropout layer is set to 0.25.
9. The method of claim 1, wherein in step five, the method for recognizing the shape and the attitude of the satellite comprises:
based on a multi-core learning (MKL) technology, feature vectors generated by C-RNN models using different convolution kernels are used as input data of a Support Vector Machine (SVM), and basic kernel functions are combined through a multi-core linear combination method, wherein the multi-core linear combination can be described as follows:
wherein X, z ∈ X,
as a result of the characteristic space,
for the ith normalized fundamental kernel function, K (x, z) represents the final kernel function formed by linearly combining n fundamental kernel functions, βiRepresents the ith coefficient; the basic kernel function is a polynomial kernel function, which can be expressed as:
in the formula: x, z ∈ X,
is a feature space; r and d are respectively constant and polynomial order; wherein, MKL is a multi-feature fusion method, and SVM is a classification model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910239623.7A CN109993224B (en) | 2019-03-27 | 2019-03-27 | GEO satellite shape and attitude identification method based on deep learning and multi-core learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910239623.7A CN109993224B (en) | 2019-03-27 | 2019-03-27 | GEO satellite shape and attitude identification method based on deep learning and multi-core learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109993224A CN109993224A (en) | 2019-07-09 |
| CN109993224B true CN109993224B (en) | 2021-02-02 |
Family
ID=67131770
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910239623.7A Active CN109993224B (en) | 2019-03-27 | 2019-03-27 | GEO satellite shape and attitude identification method based on deep learning and multi-core learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109993224B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490095A (en) * | 2019-07-31 | 2019-11-22 | 中国人民解放军战略支援部队信息工程大学 | A kind of multi-modal Fusion Features Modulation Identification method and system neural network based |
| CN110929242B (en) * | 2019-11-20 | 2020-07-10 | 上海交通大学 | Method and system for carrying out attitude-independent continuous user authentication based on wireless signals |
| CN111277564B (en) * | 2020-01-08 | 2022-06-28 | 山东浪潮科学研究院有限公司 | Enterprise network anomaly detection method and system based on dynamic storage network |
| CN111351488B (en) * | 2020-03-03 | 2022-04-19 | 南京航空航天大学 | Intelligent trajectory reconstruction reentry guidance method for aircraft |
| CN111369142B (en) * | 2020-03-04 | 2023-04-18 | 中国电子科技集团公司第五十四研究所 | Autonomous remote sensing satellite task generation method |
| CN111400754B (en) * | 2020-03-11 | 2021-10-01 | 支付宝(杭州)信息技术有限公司 | Construction method and device of user classification system for protecting user privacy |
| CN111898652A (en) * | 2020-07-10 | 2020-11-06 | 西北工业大学 | Spatial target posture classification and identification method based on convolutional neural network |
| CN113326924B (en) * | 2021-06-07 | 2022-06-14 | 太原理工大学 | Depth neural network-based key target photometric positioning method in sparse image |
| CN118409342B (en) * | 2024-07-02 | 2024-09-27 | 上海卫星互联网研究院有限公司 | Data compression method and satellite |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102800104A (en) * | 2012-06-18 | 2012-11-28 | 西安空间无线电技术研究所 | Two-dimensional scattering center automatic correlation method based on ISAR (inverse synthetic aperture radar) image sequence |
| CN103186776A (en) * | 2013-04-03 | 2013-07-03 | 西安电子科技大学 | Human detection method based on multiple features and depth information |
| CN107576949A (en) * | 2017-08-23 | 2018-01-12 | 电子科技大学 | SVDD radar target-range image recognition methods based on density weight and mixed kernel function |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101450716B (en) * | 2008-12-26 | 2010-12-29 | 中国科学院国家天文台 | Fault photo-detection method for earth synchronous transfer orbit satellite in orbit |
| CN104101297B (en) * | 2014-07-22 | 2017-02-08 | 中国科学院国家天文台 | Space object dimension acquisition method based on photoelectric observation |
| CN104570742B (en) * | 2015-01-29 | 2017-02-22 | 哈尔滨工业大学 | Feedforward PID (proportion, integration and differentiation) control based rapid high-precision relative pointing control method of noncoplanar rendezvous orbit |
| CN106022391A (en) * | 2016-05-31 | 2016-10-12 | 哈尔滨工业大学深圳研究生院 | Hyperspectral image characteristic parallel extraction and classification method |
-
2019
- 2019-03-27 CN CN201910239623.7A patent/CN109993224B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102800104A (en) * | 2012-06-18 | 2012-11-28 | 西安空间无线电技术研究所 | Two-dimensional scattering center automatic correlation method based on ISAR (inverse synthetic aperture radar) image sequence |
| CN103186776A (en) * | 2013-04-03 | 2013-07-03 | 西安电子科技大学 | Human detection method based on multiple features and depth information |
| CN107576949A (en) * | 2017-08-23 | 2018-01-12 | 电子科技大学 | SVDD radar target-range image recognition methods based on density weight and mixed kernel function |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109993224A (en) | 2019-07-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109993224B (en) | GEO satellite shape and attitude identification method based on deep learning and multi-core learning | |
| CN109961089B (en) | Small sample and zero sample image classification method based on metric learning and meta learning | |
| CN108764063B (en) | Remote sensing image time-sensitive target identification system and method based on characteristic pyramid | |
| CN108052911B (en) | Deep learning-based multi-mode remote sensing image high-level feature fusion classification method | |
| CN109214452B (en) | HRRP target identification method based on attention depth bidirectional cyclic neural network | |
| CN106845401B (en) | Pest image identification method based on multi-space convolution neural network | |
| CN113486851B (en) | Hyperspectral image classification method based on double-branch spectrum multi-scale attention network | |
| Xu et al. | RPNet: A representation learning-based star identification algorithm | |
| CN113392931B (en) | Hyperspectral open set classification method based on self-supervision learning and multitask learning | |
| CN111914929A (en) | Zero sample learning method | |
| CN107169117B (en) | Hand-drawn human motion retrieval method based on automatic encoder and DTW | |
| CN107392128A (en) | The robust image recognition methods returned based on double low-rank representations and local constraint matrix | |
| CN111652273B (en) | Deep learning-based RGB-D image classification method | |
| CN114511735A (en) | Hyperspectral image classification method and system of cascade empty spectral feature fusion and kernel extreme learning machine | |
| CN112132802A (en) | Spacecraft collision damage detection method based on self-learning splicing algorithm | |
| CN111008575A (en) | Robust face recognition method based on multi-scale context information fusion | |
| Sharma | Pose estimation of uncooperative spacecraft using monocular vision and deep learning | |
| CN114723764A (en) | Parameterized edge curve extraction method for point cloud object | |
| CN113947101A (en) | Unsupervised pedestrian re-identification method and system based on softening similarity learning | |
| CN111627007A (en) | Spacecraft defect detection method based on self-optimization matching network image stitching | |
| CN109871907B (en) | Radar target high-resolution range profile identification method based on SAE-HMM model | |
| Ozaki et al. | DNN-based self-attitude estimation by learning landscape information | |
| CN104933410A (en) | United classification method for hyper-spectral image spectrum domain and spatial domain | |
| CN113111706A (en) | SAR target feature unwrapping and identifying method for continuous missing of azimuth angle | |
| Wang et al. | Quick Recognition of Rock Images for Mobile Applications. |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |